Riddles that catch your AI being right for the wrong reasons
A right answer proves nothing. An AI can land on it by luck, by memorization, or by a shortcut that happened to point the same way. This book is the test that tells the difference.
Fifteen riddles, each with a scoring key that grades the reasoning, not the verdict. Every answer lands in one of four squares: Earned, Near miss, Lost, or Lucky. Lucky is the one to fear: the right name, with nothing behind it. Run the set on your model in an afternoon and you get four counts that say how much of its intelligence is earned.
Ten detective cases, each with a deputy in the prompt confidently naming the wrong suspect. Five structural puzzles with no story to hide in. Every riddle carries a Probe, a follow-up built so that knowing the answer doesn’t help, and four sample answers, one per square, so you learn to spot a confident paragraph that was never earned. Plus a worked example to teach the scoring hand, a kit for rebuilding riddles with fresh surfaces, and a five-minute screen for older or smaller models.
$USD 7 ยท PDF & EPUB included
WHO IT’S FOR
Anyone who delegates thinking to a model and wants to know where it guesses: developers comparing models or versions, prompt engineers checking whether standing instructions changed anything real, writers and analysts who lean on an AI every day and want to calibrate their trust.